JP3703858B2 - Attribution discrimination device - Google Patents

Attribution discrimination device Download PDF

Info

Publication number
JP3703858B2
JP3703858B2 JP24561893A JP24561893A JP3703858B2 JP 3703858 B2 JP3703858 B2 JP 3703858B2 JP 24561893 A JP24561893 A JP 24561893A JP 24561893 A JP24561893 A JP 24561893A JP 3703858 B2 JP3703858 B2 JP 3703858B2
Authority
JP
Japan
Prior art keywords
degree
group
groups
discriminant function
discrimination
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
JP24561893A
Other languages
Japanese (ja)
Other versions
JPH07105166A (en
Inventor
一之 金井
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sysmex Corp
Original Assignee
Sysmex Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sysmex Corp filed Critical Sysmex Corp
Priority to JP24561893A priority Critical patent/JP3703858B2/en
Priority to TW083107933A priority patent/TW290663B/zh
Priority to KR1019940024399A priority patent/KR100328119B1/en
Priority to US08/314,008 priority patent/US5619990A/en
Priority to DE69432885T priority patent/DE69432885D1/en
Priority to EP94402172A priority patent/EP0646881B1/en
Priority to CN94117002A priority patent/CN1099652C/en
Publication of JPH07105166A publication Critical patent/JPH07105166A/en
Application granted granted Critical
Publication of JP3703858B2 publication Critical patent/JP3703858B2/en
Anticipated expiration legal-status Critical
Expired - Fee Related legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Biomedical Technology (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Epidemiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Investigating Or Analysing Biological Materials (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Medical Treatment And Welfare Office Work (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Description

【0001】
【産業上の利用分野】
この発明は、2群線形判別の分析結果を統合して多群判別を行うことにより、被検データがどの群に帰属するのかの判別を行う帰属度判別装置に関する。この装置は、例えば疾患の診断(患者がどの疾患群に属するのかの診断)に有効に用いられる。
【0002】
【従来の技術】
2群線形判別分析は、着目する対象が2群の内のいずれの群に帰属するのかを統計的に判別する手法として、たいへんポピュラーな方法である。また、この2群線形判別分析は、知識獲得の観点からも最終的には2群線形判別関数の判別値の正負で2群の内のいずれの群に帰属するのかを判別するため、たいへん情報圧縮率が高い。
【0003】
一方、ある疾患データに対して、考えられる多群の中から1群を決定する方法(診断システム)としては、エキスパートシステムを代表とする記号化した知識と推論エンジンで演繹的に実現する方法、ニューラルネットワークの学習を利用する方法、多変量解析の応用としての多群判別のように帰納的に実現する方法などがある。
【0004】
人間たとえば熟練した専門医がどのような思考過程で多群の中から1群を決定するかを考えて見ると、決して多群から一気に1群を決定することはせず、考え得る群を数群に絞り、それらを相互に比較することにより、最も「らしい」群を決定しているのではないかと思われる。このようなプロセスは、一般に「選好プロセス」と呼ばれており、人間が比較的得意とするものである。
【0005】
このような多群パターンの認識については、電子情報通信学会論文誌 A Vol.J72-A No.1 pp.41-48(1989年1月)に、対判別による多群パターン認識に関する記載がある。
【0006】
この多群パターンの認識においては、すべての対について2群間の判別を行い、これらの結果を統合して多群への判定結果を出すようにしている。すなわち、全群に共通の分散共分散行列を仮定するのではなく、対ごとに分散共分散行列を仮定し、未知入力パターンxに対し各群Πi,Πjからのマハラノビスの距離Di2,Dj2、事後確立P(Πi|x),P(Πj|x)を計算し、マハラノビスの距離Di2,Dj2の大小、または事後確立P(Πi|x),P(Πj|x)の大小から判別結果を出し、これを基準化し(基準化対統計量)統合した上で、最終的に未知入力パターンxの分類結果を得るようにしている。ただし、事後確立P(Πi|x),P(Πj|x)に関しては、具体的な算出方法の記載はない。
【0007】
さらに、ここには、最終結果を得る方法として、多数決法、ミニマックス法、及び期待値による方法が示されている。この多数決法とは、2群判別結果を集計して多数決によって正解を得ようとするものである。ミニマックス法とは、群数が多い場合に有効であり、マハラノビスの距離のなかで最大値を求め、距離が大きくなったときその群に属することを否定するものである。期待値による方法とは、群Πiに関する基準化対統計量の期待値をとり、最良の結果を与える群を解とするものである。
【0008】
【発明が解決しようとする課題】
しかしながら、このような従来の多群判別においては、例えば、実際に各疾患を判別するシステムを構築する際には、知識獲得(線形判別関数の算出など)に用いるデータの質の違いにより、得られた線形判別関数の信憑性が違うことが普通である。
【0009】
また、実際の疾患の判別においては、非常に重要な意味を持つ組み合わせの2群判別と、あまり意味を持たない組み合わせの2群判別とがあり、すべての組み合わせの2群判別を平等に考えることは好ましくない。専門医は経験をベースとして無意識の内にこのような処理を行っていると思われる。
以上のことがらは、現実の多群判別システム、特に疾患の判別を行うシステムにとっては、たいへん重要なポイントとなる。
【0010】
この発明は、このような事情を考慮してなされたもので、2群線形判別により、被検データが多群中のどの群に帰属するのかを算出すると同時に、各2群の組み合わせの判別毎にその支持度合を求め、その支持度合を参照して被検データの帰属度を判別することにより、従来よりも高い信頼性で帰属度の判別を行えるようにした帰属度判別装置に関する。
【0011】
【課題を解決するための手段】
図1はこの発明の構成を示すブロック図であり、図に示すように、この発明は、疾患の判別に用いられる帰属度判別装置であって、多数の疾患群について各群の疾患データをあらかじめ記憶した記憶手段101と、記憶手段101に記憶された多数の疾患群の中から任意の2群を、全ての組み合わせについて選択し、選択した各2群を最適に2分する線形判別関数を、各2群毎にそれぞれ設定する判別関数設定手段102と、判別関数設定手段102によって設定された各2群毎の線形判別関数を用いて、各2群毎に被検データがどちらの群に属するのかの2群判別結果を算出する2群判別結果算出手段103と、2群判別結果算出手段103による2群判別結果の支持度合を各2群毎に決定する支持度合決定手段104と、各2群毎の2群判別結果と支持度合とに基づき、各群に対する被検データの帰属度を算出する帰属度算出手段105を有し、帰属度算出手段105が算出した各群に対する被検データの帰属度を出力装置に出力させる、帰属度判別装置である。
【0012】
この発明における記憶手段101としては、フロッピーディスク装置や磁気ディスク装置などの各種の外部記憶装置を用いることができる。
判別関数設定手段102、2群判別結果算出手段103、支持度合決定手段104、及び帰属度算出手段105としては、CPU,ROM,RAM,I/Oポートからなるマイクロコンピュータを用いることができる。
【0013】
この発明においては、支持度合決定手段104による支持度合は、僅差で判別されたのか大差で判別されたのかの度合を数量化した2群判別度合と、線形判別関数の確からしさの度合を数量化した判別関数の確信度と、判別しようとする2群の重要性の度合を数量化した重要度との、3つの要素の少なくとも1つを考慮することによって決定されることが好ましい。
【0014】
また、支持度合決定手段104の支持度合における2群判別度合は、被検データの各2群に対するマハラノビスの距離の2乗を求め、それらの数値の距離比を求め、この距離比を正規化したものを2群判別度合とすることによって求められることが好ましい。
【0015】
さらに、支持度合決定手段104の支持度合における判別関数の確信度は、判別関数設定手段によって各2群毎にそれぞれ線形判別関数が設定されるときの判別性能の評価指標の1つである相関比によって決定されることが好ましい。
【0016】
【作用】
この発明によれば、判別関数設定手段により、記憶手段に記憶された多数の群の中から任意の2群を選択し、選択した2群を最適に2分する線形判別関数を設定する。次に、その線形判別関数を用いて、2群判別結果算出手段により、被検データが2群の内どちらの群に属するのかの2群判別結果を得るとともに、支持度合決定手段により、その2群判別結果の支持度合を得、これをすべての2群対の組み合わせについて行う。
【0017】
そして、得られた2群判別結果と支持度合とに基づいて、被検データの各群に対する帰属度を算出する。このことにより、獲得知識の実状にあった的確な疾患群の判別ができる。
【0018】
【実施例】
以下、図面に示す実施例に基づいてこの発明を詳述する。なお、これによってこの発明が限定されるものではない。
【0019】
この実施例においては、血液分析装置から得られる、WBC(白血球数),RBC(赤血球数),HGB(ヘモグロビン量),MCV(平均赤血球容積),MCHC(平均赤血球ヘモグロビン濃度),PLT(血小板数),RDW(赤血球分布幅)の7種の測定データを用いて、7次元の被検データX(=X1,X2,X3,X4,X5,X6,X7)が、β−サラセミア,鉄欠乏性貧血,2次性貧血▲1▼,2次性貧血▲2▼,2次性貧血▲3▼,再生不良性貧血+MDS,溶血性貧血▲1▼,溶血性貧血▲2▼,巨赤芽球性貧血,鉄欠乏性貧血治療中,臍帯血,多血症,正常(成人),正常(小児)の14群の、いずれの群(疾患)に属するのかの判別を行う場合について説明する。
【0020】
図2はこの発明による帰属度判別装置の一実施例の構成を示すブロック図である。この図において、10はフロッピーディスク装置や磁気ディスク装置などの各種の外部記憶装置から構成されるデータベース、12は2群判別分析装置、14は例えばRAMのような記憶装置、16は制御装置、18はCRTディスプレイ装置のような表示装置やドットプリンタのような印字装置からなる出力装置である。2群判別分析装置12及び制御装置16は、CPU,ROM,RAM,I/Oポートからなるマイクロコンピュータによって構成されている。
【0021】
図3はこの発明の帰属度判別装置の処理概要を示すブロック図であり、このブロック図に基づいて本装置の処理の概要を説明する。
まず、知識獲得プロセスについて説明する。
本装置による帰属度の判別においては、上記各群(疾患)i:1,2,…,14ごとに多数の良質の知識獲得データDi:D1,D2,…,D14を確保しておくことが必要である。
【0022】
この知識獲得データ群Diは、β−サラセミア,鉄欠乏性貧血,2次性貧血▲1▼,2次性貧血▲2▼,2次性貧血▲3▼,再生不良性貧血+MDS,溶血性貧血▲1▼,溶血性貧血▲2▼,巨赤芽球性貧血,鉄欠乏性貧血治療中,臍帯血,多血症,正常(成人),及び正常(小児)の14群に所属する人の血液を実際に分析し、各群に属する人のWBC,RBC,HGB,MCV,MCHC,PLT,RDWの7種の血液測定データをそれぞれ記録したものである。この知識獲得データ群Diは、データベース10に格納されている。
【0023】
2群判別分析装置12は、データベース10からこの知識獲得データ群Diを読み出し、この知識獲得データ群Diの全ての2群対の組み合わせについて2群判別分析を行い、線形判別関数Fijなどの知識を獲得する。獲得された知識は記憶装置14に記憶される。
【0024】
次に、判別診断プロセスについて説明する。
制御装置16は、未知の被検データXを取り込み、被検データXと獲得された知識とから被検データXが各疾患に属する可能性を、2群判別結果Aij及びその支持度合Sijの数値として算出する。この被検データXとは、どの疾患に属するのかを調べようとする患者の実際のWBC,RBC,HGB,MCV,MCHC,PLT,RDWの7種の血液データである。
【0025】
数値として算出された各疾患に属する可能性の結果は、各群ごとに集計され、帰属度Kiとして出力装置18に出力される。本装置においては、未知の被検データXと上記獲得された知識とから疾患の判別を行う。
【0026】
図4は上記知識獲得プロセスと判別診断プロセスの詳細を示すブロック図である。図において、知識獲得プロセスはブロック20で示し、判別診断プロセスはブロック22で示す。以下、このブロック図を参照しながら本装置の処理について詳述する。
【0027】
2群判別分析装置12では、具体的には14群、すなわち知識獲得データD1,D2,…,D14のすべての2群対について、2群線形判別分析を行う。2群対の組み合わせ(i,j)は全部で142=(14×13)/2=91通りある。その組み合わせに対応して2群を分画するための線形判別関数Fij(X1,X2,…,X7)も91通り必要となる。
【0028】
線形判別関数Fij(X1,X2,…,X7)は次式のようになる。
Fij(X1,X2,…,X7)=aij0+aij1X1+aij2X2+…+aij7X7 式(1)
ただし、aij0,aij1,aij2,…,aij7 は係数、X1,X2,…,X7 は変数とする。
【0029】
知識獲得データの疾患群Di,Djを最良に判断する線形判別関数Fij(X1,X2,…,X7)を求める方法、すなわち、上記係数aij0,aij1aij2,…,aij7を求めるには、公知の方法を用いることができる。また、知識獲得データの疾患群Di,Djの各平均値行列Yi,Yj、各分散共分散行列Si,Sj、相関比を求める方法も公知である。これらの行列Yi,Yj,Si,Sjは、後述の被検データの群i,jへのマハラノビス距離の2乗Mi,Mjを算出する際に用いられる。
【0030】
以上の2群対(i,j)ごとに得られた線形判別関数式の係数aij0,aij1,aij2,…,aij7,平均値行列Yi,Yj、分散共分散行列Si,Sj、及び相関比が、本装置で用いる獲得された知識となる。
【0031】
制御装置16では、被検データXと獲得された知識とから被検データXが各疾患に属する可能性を数値として算出する。
すなわち、疾患の2群対(i,j)を選び、被検データXとその2群対に対応する線形判別関数Fij(X1,X2,…,X7)から、被検データXが疾患iに属するのか、疾患jに属するのかを求める。この場合、疾患iに属すると判別されたときはAij=1とし、疾患jと判別されたときはAij=0とする。この2群判別結果Aijは、被検データX:(X1,X2,…,X7)を上記式(1)に代入して得られる判別関数値の正負により決定できる。
【0032】
疾患iに属する可能性の度合Kiを算出するのに、上記2群判別結果Aijの情報だけを用いて可能性度合Kiを求めることが可能である。しかし、一般に精度はあまり良くない。なぜなら、大差で判別される場合と僅差で判別される場合があるが、その点が考慮されないからである。
【0033】
そこで、本発明では、群iと群jの2群判別結果に対する支持度合Sijの概念を導入し、同じ2群判別結果でも、どの程度の判別度合かを考慮できるようにした。2群判別結果の支持度合は、2群判別結果の程度を表すものであり、0〜1の値をとる。この支持度合が1に近いほど2群判別結果に対する支持度合が高い、つまり大差で判別されたことを示し、支持度合が0に近いほど2群判別結果に対する支持度合が低い、つまり僅差で判別されたことを示す。
【0034】
以下、具体的な2群判別結果の支持度合Sijの算出例を説明する、なお、この説明においては、判別関数の確信度Cij、重要度Wij、2群判別度合Rijに分けて、それぞれ説明する。
【0035】
(1)判別関数の確信度Cijの算出方法
判別関数の獲得は、当然ながら現実のデータから行われるため、各群の知識獲得データの質が知識である判別関数に大きく影響する。2群判別分析では、得られた2群線形判別関数Fij(X1,X2,…,X7)の判別性能を表すのに相関比を使用することがある。
【0036】
この相関比は、群iと群jの2群を判別する線形判別式に、被検データXを代入して得られる判別関数値Zの両群の度数分布を求め、各群ごとの判別関数値の平均値Yi,Yj,分散σi 2,σj 2(サンプル数:ni,nj)と、両群全体の平均値YT、分散σT 2(サンプル数:nT(=ni+nj))を算出する。
【0037】
ここで、級内分散をσW 2とすると、級内分散σW 2
σW 2=(niσi 2+njσj 2)/nT
級間分散をσB 2とすると、級間分散σB 2
σB 2=(ni(Yi−YT2+nj(Yj−YT2)/nT
となる。
【0038】
また、σT 2=σW 2+σB 2となるので、相関比は、次のように算出される。
相関比Cij=σB 2/σT 2 (0<相関比<1)
この相関比は、判別能力が高ければ級間分散が大きく、級内分散が小さくなるため、1に近い値となる。この相関比を判別関数の確信度Cijとする。
【0039】
(2)重要度Wijの設定
被検データXに対するi群への帰属度は、全14群の内、群iと群iを除く13群との2群対の線形判別の結果として求められる。群iへの帰属度を算出する際、一般的には13組(この実施例においては、1つの群に関して13組の2群対が存在する)の2群対に対して重要度に差がある場合が普通である。たとえば、鉄欠乏性貧血群への帰属度を算出する際に、β−サラセミア群、正常群などの類似群との判別は、臨床的に重要であるが、多血症群などとの判別は情報的にもあまり価値がないし、臨床的重要度も低い。このような関係を13点満点(この実施例においては、1つの群に関して13組の2群対が存在するため)で各2群対に重要度Wijとして割り当てる。
【0040】
(3)2群判別度合Rijの算出方法
被検データXの疾患群i,jへのマハラノビス距離の2乗Mi,Mjを算出する。
【0041】
疾患iへのマハラノビス距離の2乗Miは、よく知られているように、被検データX、知識獲得で得られた疾患i群の平均値行列Yi,分散共分散行列Siとから次式で求められる。
Mi=t(X−Yi)Si-1(X−Yi)
ただし、tは転置行列、-1は逆行列を表す。
疾患jへのマハラノビス距離の2乗Mjも同様にして求められる。
【0042】
2つのマハラノビス距離の2乗からMiのMjに対する相対的なマハラノビス距離の2乗比MRijを設定し、その値MRijを関数fにより0〜1の値に変換し、2群判別度合Rijとする。
【0043】
たとえば、MRij(=Mj/Mi)をxとして、
【数1】

Figure 0003703858
ただし、x≦1のときは、Rij=0とする。
で算出する。
【0044】
図5は上記関数f(x)のグラフである。この関数f(x)は正規化のための関数の一例で、1〜∞の値をとるMRijを、0〜1に連続的に変換することができる。
【0045】
定数−0.366 の根拠は、被検データXのi,j群へのマハラノビス距離の比が2のとき(i群への距離に対して、j群への距離が2倍のとき:MRijは2乗であるので4とする)に、2群判別度合Rijが0.5となるように設定されている。この定数−0.366 に関しては、判別する群の相対的な近さにより変更する必要がある。
なお、他に、知識獲得データに2群線形判別式を適用したときの判別的中率を用いる方法もある。
【0046】
以上の結果より、判別結果に対する支持度合Sijを
Sij=Rij×Cij×Wij
で求め、次式でi群に対する帰属度Kiを算出する。
【数2】
Figure 0003703858
ただし、n=14,j≠iとする。また、Aij=1の場合は疾患i側に判別されたとき、Aij=0の場合は疾患j側に判別されたときである。
上記の帰属度Kiは、0≦Ki≦1となる。
【0047】
支持度合Sijは、2群判別度合Rijだけでも従来よりも判別精度を良くすることができるが、判別関数の確信度Cijも考慮すると(Sij=Rij×Cij)、知識獲得データの実状にあったものとなる。
そして、さらに重要度Wijまで考慮すると(Sij=Rij×Cij×Wij)、より一層人間の判断に近づけることが可能である。
【0048】
次に、このような処理動作の内容を図6及び図7に示すフローチャートに基づいて説明する。
図6は知識獲得プロセスと判別診断プロセスの全体の処理内容を示すフローチャートである。
【0049】
この図に示すように、本装置の処理では、まず、知識獲得プロセスにおいて、知識獲得データ群Diから、各2群対ごとに、線形判別関数Fij,平均値行列Yi,分散共分散行列Si,相関比Cijを獲得する(ステップ31)。
【0050】
次に、各2群対ごとの重要度Wijを設定する(ステップ32)。なお、この重要度Wijは、知識獲得データ群Diと共に、あらかじめ設定しておいてもよい。
【0051】
そして、判別診断プロセスにおいて、これらの線形判別関数Fij,平均値行列Yi,分散共分散行列Si,相関比Cij,重要度Wijを用いて、被検データXに対して多群判別を行う(ステップ33)。
【0052】
図7は判別診断プロセスの詳細内容を示すフローチャートである。
判別診断プロセスの処理においては、まず、被検データXが入力されると(ステップ41)、任意の群iを指定し(ステップ42)、被検データXを線形判別関数式Fijに代入し、2群判別結果Aijを決定する(ステップ43)。
【0053】
続いて、被検データXと、平均値行列Yi及び分散共分散行列Siとから、マハラノビスの距離Mi,Mjを算出し(ステップ44)、マハラノビスの距離Mi,Mjの2乗比MRijに基づいて、2群判別度合Rijを算出する(ステップ45)。
次に、相関比Cij,重要度Wij,2群判別度合Rijにより、2群判別結果Aijの支持度合Sijを算出する(ステップ46)。
【0054】
ここで、指定群iについての全組み合わせij(ただし、j≠i)が終了したか否かを調べ(ステップ47)、終了すれば、2群判別結果Aijとその2群判別結果の支持度合Sijとから、指定群iの帰属度Kiを算出する(ステップ48)。
【0055】
そして、すべての群iについて終了したのか否かを調べ(ステップ49)、すべての群iについて終了すれば、帰属度Kiを値の高い順に並び換え(ステップ50)、最も可能性の高い群iを決定する(ステップ51)。
【0056】
以上の処理を実際に行った場合の出力帳票の例を図8〜図10に示す。これらの図に示したものは、2次性貧血▲3▼の疾患を有する検体を被検データとし、上記のプロセスを経て各群の帰属度Kiを求めたものである。
【0057】
これらの図の内、図8は2群判別結果だけを用いて2群判別結果の支持度合Sijを考慮しない場合(支持モード:なし)の例を示し、図9は2群判別結果の支持度合Sijの内、2群判別度合Rijだけを考慮した場合(支持モード:距離比)を示し、図10は2群判別結果の支持度合Sijの内、2群判別度合Rijと判別関数の確信度Cijを考慮した場合(支持モード:距離比,相関比)を示している。
【0058】
各モードにおける帰属度Kiの算出式は次のようになる。
支持モード:なしの場合
【数3】
Figure 0003703858
支持モード:距離比の場合
【数4】
Figure 0003703858
支持モード:距離比,相関比の場合
【数5】
Figure 0003703858
【0059】
判別結果の支持度合Sijを考慮しない場合には、図8に示すように、2次性貧血▲3▼群と溶血性貧血▲1▼の帰属度Kiが同じであったが、判別結果の支持度合Sijを考慮した場合には、図9の2群判別度合Rijだけを考慮した場合、及び図10の2群判別度合Rijと判別関数の確信度Cijとの双方を考慮した場合ともに、2次性貧血▲3▼の帰属度Kiが第1位となっている。
【0060】
さらに、支持度合Sijについて、判別関数の確信度Cijを考慮しない場合(図9)と考慮した場合(図10)とを比べると、考慮した場合(図10)の方が1位、2位の差が相対的に大きくなっている。
【0061】
このようにして、知識獲得データ群D1,…,Dnから、すべての2群対の組み合わせに対して2群判別分析を行い、知識となる判別関数Fijを得る。そして、この判別関数Fijを用いて未知の被検データXについて全組み合わせの2群判別分析を行い、2群判別結果Aij、支持度合Sijを得、この2群判別結果Aijと支持度合Sijとから被検データXの各群に対する帰属度Kiを算出することにより、獲得知識の実状にあった適確な疾患群の判別を行うことができる。
【0062】
【発明の効果】
この発明によれば、多数の疾患群の中から任意の2群を選択し、選択した2群を最適に2分する線形判別関数を設定し、その線形判別関数を用いて被検データが2群の内どちらの群に属するのかの2群判別結果を得るとともに、その2群判別結果の支持度合を得、得られた2群判別結果と支持度合とに基づいて、被検データの各群に対する帰属度を算出するようにしたので、獲得知識の実状にあった的確な疾患群の判別が可能となる。
【図面の簡単な説明】
【図1】この発明の構成を示すブロック図である。
【図2】この発明による帰属度判別装置の一実施例の構成を示すブロック図である。
【図3】この発明の帰属度判別装置の処理概要を示すブロック図である。
【図4】知識獲得プロセスと判別診断プロセスの詳細を示すブロック図である。
【図5】2群判別度合Rijの関数f(x)を示すグラフである。
【図6】知識獲得プロセスと判別診断プロセスの全体の処理内容を示すフローチャートである。
【図7】判別診断プロセスの詳細内容を示すフローチャートである。
【図8】実際に帰属度の判別を行った場合の出力帳票(支持モード:なし)の例を示す説明図である。
【図9】実際に帰属度の判別を行った場合の出力帳票(支持モード:距離比)の例を示す説明図である。
【図10】実際に帰属度の判別を行った場合の出力帳票(支持モード:距離比,相関比)の例を示す説明図である。
【符号の説明】
10 データベース
12 2群判別分析装置
14 記憶装置
16 制御装置
18 出力装置[0001]
[Industrial application fields]
The present invention relates to an belonging degree discriminating apparatus that discriminates to which group test data belongs by integrating the analysis results of the two-group linear discrimination and performing multi-group discrimination. This device is effectively used, for example, for diagnosis of a disease (diagnosis of which disease group a patient belongs to).
[0002]
[Prior art]
The two-group linear discriminant analysis is a very popular method as a method of statistically discriminating which group of the two groups the subject of interest belongs to. In addition, from the viewpoint of knowledge acquisition, this two-group linear discriminant analysis ultimately determines which group of the two groups belongs to the positive or negative of the discriminant value of the two-group linear discriminant function. High compression ratio.
[0003]
On the other hand, as a method (diagnostic system) for determining one group out of many possible groups for certain disease data, a method of a priori implementation with symbolized knowledge represented by an expert system and an inference engine, There are a method of using learning of a neural network and a method of inductive realization such as multigroup discrimination as an application of multivariate analysis.
[0004]
Considering what kind of thinking process a human, for example, a skilled specialist decides one group out of many groups, never decides one group at a time from many groups, but several groups of possible groups It seems that the most “probable” group is determined by narrowing down to and comparing them with each other. Such a process is generally called a “preference process” and is relatively good at humans.
[0005]
Regarding such multi-group pattern recognition, the IEICE Transactions A Vol.J72-A No.1 pp.41-48 (January 1989) describes multi-group pattern recognition by pair discrimination. .
[0006]
In recognition of this multi-group pattern, discrimination between two groups is performed for all pairs, and these results are integrated to obtain a determination result for a multi-group. That is, instead of assuming a covariance matrix common to all groups, a covariance matrix is assumed for each pair, and Mahalanobis distances Di 2 and Dj 2 from each group Πi and Πj with respect to the unknown input pattern x. Then, post-establishment P (Πi | x), P (Πj | x) is calculated, and the Mahalanobis distances Di 2 and Dj 2 are calculated from the magnitudes of the post-establishment P (Πi | x) and P (Πj | x). A discrimination result is output, standardized (standardized versus statistics), integrated, and finally a classification result of the unknown input pattern x is obtained. However, no specific calculation method is described for the post-establishment P (確立 i | x) and P (Πj | x).
[0007]
Furthermore, here, as a method for obtaining a final result, a majority method, a minimax method, and a method based on an expected value are shown. The majority method is to collect the two-group discrimination results and try to obtain a correct answer by majority vote. The minimax method is effective when the number of groups is large, and the maximum value is obtained from the Mahalanobis distance, and when the distance increases, it is denied that the group belongs to that group. The expected value method is a method that takes the expected value of the standardization versus statistics for the group Πi and sets the group that gives the best result as the solution.
[0008]
[Problems to be solved by the invention]
However, in such conventional multi-group discrimination, for example, when a system for actually discriminating each disease is constructed, it is obtained due to the difference in data quality used for knowledge acquisition (calculation of linear discriminant function, etc.). It is normal that the credibility of the given linear discriminant function is different.
[0009]
In actual disease discrimination, there are two-group discrimination for combinations with very important meanings and two-group discrimination for combinations with little meaning, and consider two-group discrimination for all combinations equally. Is not preferred. The specialist seems to be doing such a process unconsciously based on experience.
The above is a very important point for an actual multi-group discrimination system, particularly a system for discriminating diseases.
[0010]
The present invention has been made in consideration of such circumstances. By calculating the group to which the test data belongs in the multi-group by the 2-group linear discrimination, at the same time, for each discrimination of the combination of each 2 groups. Further, the present invention relates to a degree-of-attachment discriminating apparatus which can determine the degree of belonging with higher reliability than before by determining the degree of support and determining the degree of belonging of test data with reference to the degree of support.
[0011]
[Means for Solving the Problems]
FIG. 1 is a block diagram showing the configuration of the present invention. As shown in the figure, the present invention is an attribution degree discriminating apparatus used for disease discrimination, in which disease data of each group is preliminarily stored for a number of disease groups. Stored storage means 101, and a linear discriminant function that selects any two groups from among a large number of disease groups stored in storage means 101 for all combinations, and optimally bisects each selected two groups, Using the discriminant function setting means 102 set for each two groups and the linear discriminant function for each two groups set by the discriminant function setting means 102, the test data belongs to which group for each two groups 2 group discrimination result calculation means 103 for calculating the 2 group discrimination result of No. 2; support degree determination means 104 for determining the support degree of the 2 group discrimination result by the 2 group discrimination result calculation means 103 for each 2 groups; 2 groups per group Based on the different results and support degree, it has a degree of membership calculation means 105 for calculating the degree of membership of the test data for each group, output devices attribution degree of the test data for each group were calculated attribution degree calculating unit 105 It is an attribution degree discriminating device that outputs to
[0012]
As the storage means 101 in the present invention, various external storage devices such as a floppy disk device and a magnetic disk device can be used.
As the discrimination function setting means 102, the 2-group discrimination result calculation means 103, the support level determination means 104, and the attribution level calculation means 105, a microcomputer comprising a CPU, a ROM, a RAM, and an I / O port can be used.
[0013]
In the present invention, the degree of support by the degree-of-support determining means 104 is quantified as a two-group discriminating degree obtained by quantifying the degree of whether it is discriminated with a small difference or with a large difference, and the degree of probability of the linear discriminant function. Preferably, it is determined by considering at least one of the three factors of the certainty of the discriminant function and the importance obtained by quantifying the importance of the two groups to be discriminated.
[0014]
Further, the two-group discrimination degree in the support degree of the support degree determining means 104 is obtained by calculating the square of the Mahalanobis distance to each two groups of the test data, obtaining a distance ratio of those numerical values, and normalizing the distance ratio. It is preferable that it is calculated | required by making a thing into 2 group discrimination | determination degree.
[0015]
Further, the certainty factor of the discriminant function in the support level of the support level determining unit 104 is a correlation ratio that is one of the evaluation indexes of the discrimination performance when the linear discriminant function is set for each of the two groups by the discriminant function setting unit. Is preferably determined by:
[0016]
[Action]
According to the present invention, the discriminant function setting means selects two arbitrary groups from among a large number of groups stored in the storage means, and sets a linear discriminant function that optimally bisects the selected two groups. Next, using the linear discriminant function, the second group discriminant result calculating means obtains the second group discriminant result of which of the two groups the test data belongs, and the support degree determining means 2 The degree of support of the group discrimination result is obtained, and this is performed for all combinations of two groups.
[0017]
Then, based on the obtained two-group discrimination result and the degree of support, the degree of belonging to each group of the test data is calculated. This makes it possible to accurately identify a disease group that matches the actual state of acquired knowledge.
[0018]
【Example】
Hereinafter, the present invention will be described in detail based on embodiments shown in the drawings. However, this does not limit the present invention.
[0019]
In this example, WBC (white blood cell count), RBC (red blood cell count), HGB (hemoglobin content), MCV (average red blood cell volume), MCHC (average red blood cell hemoglobin concentration), PLT (platelet count) obtained from the blood analyzer. ), RDW (red blood cell distribution width) using 7 kinds of measurement data, 7-dimensional test data X (= X1, X2, X3, X4, X5, X6, X7) is β-thalassemia, iron deficiency Anemia, secondary anemia (1), secondary anemia (2), secondary anemia (3), aplastic anemia + MDS, hemolytic anemia (1), hemolytic anemia (2), giant erythroblast A case will be described in which a group (disease) belonging to 14 groups of umbilical cord blood, polycythemia, normal (adult), and normal (child) is determined during treatment of anemia and iron deficiency anemia.
[0020]
FIG. 2 is a block diagram showing a configuration of an embodiment of the belonging degree discriminating apparatus according to the present invention. In this figure, 10 is a database composed of various external storage devices such as a floppy disk device and a magnetic disk device, 12 is a two-group discriminant analysis device, 14 is a storage device such as a RAM, 16 is a control device, 18 Is an output device comprising a display device such as a CRT display device or a printing device such as a dot printer. The two-group discriminant analysis device 12 and the control device 16 are constituted by a microcomputer comprising a CPU, a ROM, a RAM, and an I / O port.
[0021]
FIG. 3 is a block diagram showing an outline of the process of the belonging degree discriminating apparatus of the present invention. The outline of the process of the present apparatus will be described based on this block diagram.
First, the knowledge acquisition process will be described.
In the determination of the degree of attribution by this apparatus, it is necessary to secure a large number of high-quality knowledge acquisition data Di: D1, D2,..., D14 for each group (disease) i: 1, 2,. is necessary.
[0022]
This knowledge acquisition data group Di includes β-thalassemia, iron deficiency anemia, secondary anemia (1), secondary anemia (2), secondary anemia (3), aplastic anemia + MDS, hemolytic anemia (1), hemolytic anemia (2), megaloblastic anemia, iron deficiency anemia, umbilical cord blood, polycythemia, normal (adult) and normal (children) belonging to 14 groups The blood is actually analyzed, and seven types of blood measurement data of WBC, RBC, HGB, MCV, MCHC, PLT, and RDW of each person belonging to each group are recorded. This knowledge acquisition data group Di is stored in the database 10.
[0023]
The two-group discriminant analyzer 12 reads out this knowledge acquisition data group Di from the database 10, performs a two-group discriminant analysis on all combinations of two groups of the knowledge acquisition data group Di, and acquires knowledge such as the linear discriminant function Fij. To win. The acquired knowledge is stored in the storage device 14.
[0024]
Next, the discrimination diagnosis process will be described.
The control device 16 takes in the unknown test data X, and determines the possibility that the test data X belongs to each disease from the test data X and the acquired knowledge, and is a numerical value of the second group discrimination result Aij and its support level Sij. Calculate as The test data X is seven kinds of blood data of actual WBC, RBC, HGB, MCV, MCHC, PLT, and RDW of a patient to be examined for which disease.
[0025]
The result of possibility of belonging to each disease calculated as a numerical value is totaled for each group, and is output to the output device 18 as the degree of belonging Ki. In this apparatus, the disease is discriminated from the unknown test data X and the acquired knowledge.
[0026]
FIG. 4 is a block diagram showing details of the knowledge acquisition process and the discrimination diagnosis process. In the figure, the knowledge acquisition process is indicated by block 20 and the discriminative diagnosis process is indicated by block 22. Hereinafter, the processing of this apparatus will be described in detail with reference to this block diagram.
[0027]
Specifically, the two-group discriminant analyzer 12 performs the two-group linear discriminant analysis on all 14 groups, that is, all 2 group pairs of the knowledge acquisition data D1, D2,. There are a total of 14 C 2 = (14 × 13) / 2 = 91 combinations of the two groups of pairs (i, j). 91 kinds of linear discriminant functions Fij (X1, X2,..., X7) for fractionating the two groups corresponding to the combinations are also required.
[0028]
The linear discriminant function Fij (X1, X2,..., X7) is as follows.
Fij (X1, X2, ..., X7) = aij 0 + aij 1 X1 + aij 2 X2 + ... + aij 7 X7 Equation (1)
Here, aij 0 , aij 1 , aij 2 ,..., Aij 7 are coefficients, and X1, X2,.
[0029]
Linear discriminant function Fij (X1, X2, ..., X7) for determining disease group Di of knowledge acquisition data, a Dj best method of obtaining a, i.e., the coefficients aij 0, aij 1 aij 2, ..., to determine the aij 7 A known method can be used. Also, a method for obtaining each average value matrix Yi, Yj, each variance-covariance matrix Si, Sj, and correlation ratio of the disease groups Di, Dj of the knowledge acquisition data is also known. These matrices Yi, Yj, Si, Sj are used when calculating squares Mi, Mj of Mahalanobis distances to groups i, j of test data described later.
[0030]
The coefficients aij 0 , aij 1 , aij 2 ,..., Aij 7 of the linear discriminant function equations obtained for each of the above two group pairs (i, j), mean value matrices Yi, Yj, variance-covariance matrices Si, Sj, And the correlation ratio is the acquired knowledge used in the present apparatus.
[0031]
The control device 16 calculates the possibility that the test data X belongs to each disease from the test data X and the acquired knowledge as a numerical value.
That is, two groups of diseases (i, j) are selected, and from the test data X and the linear discriminant function Fij (X1, X2,..., X7) corresponding to the two groups of pairs, the test data X becomes the disease i. It is determined whether it belongs to the disease j. In this case, Aij = 1 when it is determined to belong to the disease i, and Aij = 0 when it is determined to be the disease j. This two-group discrimination result Aij can be determined by the sign of the discriminant function value obtained by substituting the test data X: (X1, X2,..., X7) into the above equation (1).
[0032]
In order to calculate the degree Ki of the possibility of belonging to the disease i, it is possible to obtain the possibility degree Ki using only the information of the second group discrimination result Aij. However, accuracy is generally not very good. This is because there is a case where the discrimination is made with a large difference and a case where the discrimination is made with a slight difference, but this point is not taken into consideration.
[0033]
Therefore, in the present invention, the concept of the degree of support Sij for the two-group discrimination result of the group i and the group j is introduced so that the degree of discrimination can be considered even with the same two-group discrimination result. The support level of the second group discrimination result represents the degree of the second group discrimination result, and takes a value of 0 to 1. The closer the support level is to 1, the higher the support level for the second group discrimination result is, that is, it is discriminated by a large difference. The closer the support level is to 0, the lower the support level for the second group discrimination result is, that is, the discriminant is a little difference. It shows that.
[0034]
Hereinafter, a specific example of calculating the support level Sij of the second group discrimination result will be described. In this description, the discriminant function confidence level Cij, the importance level Wij, and the second group discrimination level Rij will be described separately. .
[0035]
(1) Method of calculating discriminant function confidence Cij Since the acquisition of the discriminant function is naturally performed from the actual data, the quality of the knowledge acquisition data of each group greatly affects the discriminant function that is knowledge. In the two-group discriminant analysis, a correlation ratio may be used to represent the discrimination performance of the obtained two-group linear discriminant function Fij (X1, X2,..., X7).
[0036]
This correlation ratio is obtained by obtaining the frequency distribution of both groups of the discriminant function value Z obtained by substituting the test data X into the linear discriminant for discriminating between the two groups i and j. Average values Yi, Yj, variances σ i 2 , σ j 2 (number of samples: ni, nj) and average values Y T and variances σ T 2 (number of samples: n T (= n i + n) of both groups. j ))) is calculated.
[0037]
Here, when the intra-class variance is σ W 2 , the intra-class variance σ W 2 is σ W 2 = (n i σ i 2 + n j σ j 2 ) / n T
When the inter-class variance is σ B 2 , the inter-class variance σ B 2 is σ B 2 = (n i (Y i −Y T ) 2 + n j (Y j −Y T ) 2 ) / n T
It becomes.
[0038]
Since σ T 2 = σ W 2 + σ B 2 , the correlation ratio is calculated as follows.
Correlation ratio Cij = σ B 2 / σ T 2 (0 <correlation ratio <1)
This correlation ratio is close to 1 because the interclass variance is large and the intraclass variance is small if the discrimination ability is high. This correlation ratio is defined as a certainty factor Cij of the discriminant function.
[0039]
(2) Setting of Importance Wij The degree of belonging to group i with respect to test data X is obtained as a result of two-group linear discrimination between group i and group 13 excluding group i out of all 14 groups. When calculating the degree of belonging to group i, there is generally a difference in importance for two groups of 13 groups (in this example, there are 13 groups of 2 group pairs for one group). Some cases are common. For example, when calculating the degree of belonging to an iron deficiency anemia group, discrimination from similar groups such as β-thalassemia group and normal group is clinically important. It is not very informative and has little clinical importance. Such a relationship is assigned as importance Wij to each of the two group pairs with a full score of 13 (in this embodiment, because there are 13 pairs of two group pairs for one group).
[0040]
(3) Method for calculating the 2-group discrimination degree Rij The squares Mi and Mj of the Mahalanobis distance to the disease groups i and j in the test data X are calculated.
[0041]
As is well known, the square Mi of the Mahalanobis distance to the disease i is expressed by the following equation from the test data X, the mean value matrix Yi of the disease i group obtained by knowledge acquisition, and the variance-covariance matrix Si. Desired.
Mi = t (X−Yi) Si −1 (X−Yi)
Where t is a transposed matrix and -1 is an inverse matrix.
The square Mj of the Mahalanobis distance to the disease j is obtained in the same manner.
[0042]
A square ratio MRij of the Mahalanobis distance relative to Mj from the square of the two Mahalanobis distances is set, and the value MRij is converted into a value of 0 to 1 by the function f to obtain the second group discrimination degree Rij.
[0043]
For example, let MRij (= Mj / Mi) be x,
[Expression 1]
Figure 0003703858
However, when x ≦ 1, Rij = 0.
Calculate with
[0044]
FIG. 5 is a graph of the function f (x). This function f (x) is an example of a function for normalization, and MRij having a value of 1 to ∞ can be continuously converted to 0 to 1.
[0045]
The basis of the constant −0.366 is that when the ratio of the Mahalanobis distance to the i and j groups of the test data X is 2 (when the distance to the j group is twice the distance to the i group: MRij is 2 It is set so that the second group discrimination degree Rij is 0.5. This constant -0.366 needs to be changed depending on the relative proximity of the group to be discriminated.
In addition, there is another method that uses a discriminant probability when a two-group linear discriminant is applied to knowledge acquisition data.
[0046]
From the above results, the support degree Sij for the discrimination result is expressed as Sij = Rij × Cij × Wij.
And the degree of attribution Ki for the i group is calculated by the following equation.
[Expression 2]
Figure 0003703858
However, n = 14 and j ≠ i. Further, when Aij = 1, it is determined that the disease is on the disease i side, and when Aij = 0 is determined when it is determined on the disease j side.
The degree of attribution Ki is 0 ≦ Ki ≦ 1.
[0047]
The support degree Sij can improve the discrimination accuracy compared to the conventional case only with the two-group discrimination degree Rij, but considering the certainty Cij of the discrimination function (Sij = Rij × Cij), it was in the actual state of knowledge acquisition data. It will be a thing.
Further, when considering the importance Wij (Sij = Rij × Cij × Wij), it is possible to make it closer to human judgment.
[0048]
Next, the contents of such processing operation will be described based on the flowcharts shown in FIGS.
FIG. 6 is a flowchart showing the entire processing contents of the knowledge acquisition process and the discrimination diagnosis process.
[0049]
As shown in this figure, in the processing of this apparatus, first, in the knowledge acquisition process, from the knowledge acquisition data group Di, for each two group pairs, linear discriminant function Fij, mean value matrix Yi, variance covariance matrix Si, A correlation ratio Cij is acquired (step 31).
[0050]
Next, the importance Wij for each pair of two groups is set (step 32). The importance Wij may be set in advance together with the knowledge acquisition data group Di.
[0051]
In the discriminant diagnosis process, multi-group discrimination is performed on the test data X using the linear discriminant function Fij, the mean value matrix Yi, the variance-covariance matrix Si, the correlation ratio Cij, and the importance Wij (step) 33).
[0052]
FIG. 7 is a flowchart showing the detailed contents of the discrimination diagnosis process.
In the process of the discriminant diagnosis process, first, when test data X is input (step 41), an arbitrary group i is designated (step 42), and the test data X is substituted into the linear discriminant function expression Fij. The second group discrimination result Aij is determined (step 43).
[0053]
Subsequently, Mahalanobis distances Mi and Mj are calculated from the test data X, the mean value matrix Yi and the variance-covariance matrix Si (step 44), and based on the square ratio MRij of the Mahalanobis distances Mi and Mj. The second group discrimination degree Rij is calculated (step 45).
Next, the support degree Sij of the second group discrimination result Aij is calculated from the correlation ratio Cij, the importance degree Wij, and the second group discrimination degree Rij (step 46).
[0054]
Here, it is checked whether or not all combinations ij (where j ≠ i) for the designated group i have been completed (step 47), and if completed, the second group discrimination result Aij and the support level Sij of the second group discrimination result. From the above, the degree of attribution Ki of the designated group i is calculated (step 48).
[0055]
Then, it is checked whether or not all groups i have been completed (step 49). If all groups i have been completed, the belonging degree Ki is rearranged in descending order (step 50), and the most likely group i is determined. Is determined (step 51).
[0056]
Examples of output forms when the above processing is actually performed are shown in FIGS. In these figures, specimens having a disease of secondary anemia (3) are used as test data, and the degree of membership Ki of each group is obtained through the above process.
[0057]
Of these figures, FIG. 8 shows an example in which the support level Sij of the second group discrimination result is not considered using only the second group discrimination result (support mode: none), and FIG. 9 shows the support level of the second group discrimination result. FIG. 10 shows a case where only the second group discrimination degree Rij is considered (support mode: distance ratio) in Sij, and FIG. 10 shows the second group discrimination degree Rij and the discriminant function confidence Cij among the support degrees Sij of the second group discrimination result. (Support mode: distance ratio, correlation ratio) is shown.
[0058]
The calculation formula of the degree of membership Ki in each mode is as follows.
Support mode: None [Formula 3]
Figure 0003703858
Support mode: Distance ratio [Equation 4]
Figure 0003703858
Support mode: Distance ratio, correlation ratio [Equation 5]
Figure 0003703858
[0059]
When the support level Sij of the discrimination result is not taken into consideration, as shown in FIG. 8, the attribution degree Ki of the secondary anemia <3> group and the hemolytic anemia <1> is the same. When considering the degree Sij, both the case where only the second group discrimination degree Rij in FIG. 9 is considered and the case where both the second group discrimination degree Rij and the certainty factor Cij of the discriminant function are taken into consideration are secondary. The degree of attribution Ki of sexual anemia (3) is first.
[0060]
Further, regarding the support degree Sij, when the certainty factor Cij of the discriminant function is not considered (FIG. 9) and the case (FIG. 10) is compared, the case (FIG. 10) that is considered is first and second. The difference is relatively large.
[0061]
In this way, from the knowledge acquisition data groups D1,..., Dn, the two-group discriminant analysis is performed on all combinations of two groups, and the discriminant function Fij serving as knowledge is obtained. Then, using this discriminant function Fij, the two-group discriminant analysis of all combinations is performed on the unknown test data X to obtain the second group discriminant result Aij and the support degree Sij. By calculating the degree of membership Ki for each group of the test data X, it is possible to determine an appropriate disease group in accordance with the actual state of acquired knowledge.
[0062]
【The invention's effect】
According to this invention, two arbitrary groups are selected from a large number of disease groups, a linear discriminant function that optimally bisects the selected two groups is set, and test data is 2 using the linear discriminant function. Obtaining the two-group discrimination result as to which of the groups belongs, obtaining the support level of the two-group discrimination result, and based on the obtained two-group discrimination result and the support level, each group of the test data Since the degree of belonging to is calculated, it is possible to accurately determine a disease group that matches the actual state of acquired knowledge.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of the present invention.
FIG. 2 is a block diagram showing a configuration of an embodiment of an belonging degree discriminating apparatus according to the present invention.
FIG. 3 is a block diagram showing an outline of processing of the belonging degree discriminating apparatus of the present invention.
FIG. 4 is a block diagram showing details of a knowledge acquisition process and a discrimination diagnosis process.
FIG. 5 is a graph showing a function f (x) of the second group discrimination degree Rij.
FIG. 6 is a flowchart showing the entire processing contents of a knowledge acquisition process and a discrimination diagnosis process.
FIG. 7 is a flowchart showing detailed contents of a discrimination diagnosis process.
FIG. 8 is an explanatory diagram showing an example of an output form (support mode: none) when the attribution level is actually determined.
FIG. 9 is an explanatory diagram showing an example of an output form (support mode: distance ratio) when the belonging degree is actually determined.
FIG. 10 is an explanatory diagram showing an example of an output form (support mode: distance ratio, correlation ratio) when the attribution level is actually determined.
[Explanation of symbols]
10 Database 12 Group 2 Discriminant Analysis Device 14 Storage Device 16 Control Device 18 Output Device

Claims (5)

疾患の判別に用いられる帰属度判別装置であって、
多数の疾患群について各群の疾患データをあらかじめ記憶した記憶手段と、
記憶手段に記憶された多数の疾患群の中から任意の2群を、全ての組み合わせについて選択し、選択した各2群を最適に2分する線形判別関数を、各2群毎にそれぞれ設定する判別関数設定手段と、
判別関数設定手段によって設定された各2群毎の線形判別関数を用いて、各2群毎に被検データがどちらの群に属するのかの2群判別結果を算出する2群判別結果算出手段と、
2群判別結果算出手段による2群判別結果の支持度合を各2群毎に決定する支持度合決定手段と、
各2群毎の2群判別結果と支持度合とに基づき、各群に対する被検データの帰属度を算出する帰属度算出手段を有し、
帰属度算出手段が算出した各群に対する被検データの帰属度を出力装置に出力させる、帰属度判別装置。
An apparatus for discriminating the degree of belonging used for discrimination of a disease,
Storage means for storing disease data of each group in advance for a number of disease groups;
Arbitrary two groups are selected for all combinations from among a large number of disease groups stored in the storage means, and a linear discriminant function that optimally bisects each selected two group is set for each of the two groups. Discriminant function setting means;
A two-group discrimination result calculation means for calculating a two-group discrimination result to which group the test data belongs for each two groups using a linear discriminant function for each two groups set by the discrimination function setting means; ,
A support degree determining means for determining the support degree of the second group discrimination result by the second group discrimination result calculating means for each two groups;
Based on the two-group discrimination result and the degree of support for each two groups, it has an attribution calculation means for calculating the attribution of the test data for each group ,
An attribution degree discriminating apparatus for causing the output apparatus to output the attribution degree of the test data for each group calculated by the attribution degree calculating means .
支持度合決定手段による支持度合が、僅差で判別されたのか大差で判別されたのかの度合を数量化した2群判別度合と、線形判別関数の確からしさの度合を数量化した判別関数の確信度と、判別しようとする2群の重要性の度合を数量化した重要度との、3つの要素の少なくとも1つを考慮することによって決定されることを特徴とする請求項1記載の帰属度判別装置。  2-group discriminating degree quantifying the degree of whether the supporting degree by the supporting degree determining means is discriminated by a small difference or a large difference, and the certainty of the discriminant function quantifying the degree of probability of the linear discriminant function 2. The degree-of-association determination according to claim 1, wherein the determination is made by considering at least one of three elements: an importance obtained by quantifying the importance of the two groups to be determined. apparatus. 支持度合決定手段の支持度合における2群判別度合が、被検データの各2群に対するマハラノビスの距離の2乗を求め、それらの数値の距離比を求め、この距離比を正規化したものを2群判別度合とすることによって求められることを特徴とする請求項2記載の帰属度判別装置。  The two-group discrimination degree in the support degree of the support degree determining means obtains the square of the Mahalanobis distance to each two groups of the test data, obtains a distance ratio of those numerical values, and normalizes this distance ratio to 2 3. The belonging degree discriminating apparatus according to claim 2, wherein the belonging degree discriminating apparatus is obtained by using a group discriminating degree. 支持度合決定手段の支持度合における判別関数の確信度が、判別関数設定手段によって各2群毎にそれぞれ線形判別関数が設定されるときの判別性能の評価指標の1つである相関比によって決定されることを特徴とする請求項2記載の帰属度判別装置。  The certainty of the discriminant function in the support degree of the support degree determining means is determined by the correlation ratio which is one of the evaluation indexes of the discriminating performance when the linear discriminant function is set for each of the two groups by the discriminant function setting means. The belonging degree discriminating apparatus according to claim 2. 前記被検データが、血液分析装置から得られる血液データである請求項1記載の帰属度判別装置。  2. The belonging degree discriminating apparatus according to claim 1, wherein the test data is blood data obtained from a blood analyzer.
JP24561893A 1993-09-30 1993-09-30 Attribution discrimination device Expired - Fee Related JP3703858B2 (en)

Priority Applications (7)

Application Number Priority Date Filing Date Title
JP24561893A JP3703858B2 (en) 1993-09-30 1993-09-30 Attribution discrimination device
TW083107933A TW290663B (en) 1993-09-30 1994-08-30
US08/314,008 US5619990A (en) 1993-09-30 1994-09-28 Apparatus and method for making a medical diagnosis by discriminating attribution degrees
KR1019940024399A KR100328119B1 (en) 1993-09-30 1994-09-28 Judging device
DE69432885T DE69432885D1 (en) 1993-09-30 1994-09-29 Device and method for supporting a medical diagnosis by distinguishing degrees of assignment
EP94402172A EP0646881B1 (en) 1993-09-30 1994-09-29 Apparatus and method for assisting a medical diagnosis by discriminating attribution degrees
CN94117002A CN1099652C (en) 1993-09-30 1994-09-30 Belongingness judging device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
JP24561893A JP3703858B2 (en) 1993-09-30 1993-09-30 Attribution discrimination device

Publications (2)

Publication Number Publication Date
JPH07105166A JPH07105166A (en) 1995-04-21
JP3703858B2 true JP3703858B2 (en) 2005-10-05

Family

ID=17136376

Family Applications (1)

Application Number Title Priority Date Filing Date
JP24561893A Expired - Fee Related JP3703858B2 (en) 1993-09-30 1993-09-30 Attribution discrimination device

Country Status (7)

Country Link
US (1) US5619990A (en)
EP (1) EP0646881B1 (en)
JP (1) JP3703858B2 (en)
KR (1) KR100328119B1 (en)
CN (1) CN1099652C (en)
DE (1) DE69432885D1 (en)
TW (1) TW290663B (en)

Families Citing this family (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5797839A (en) * 1996-12-23 1998-08-25 Apple Blossom, Llc Homeopathic treatment selection system and method
US5860917A (en) * 1997-01-15 1999-01-19 Chiron Corporation Method and apparatus for predicting therapeutic outcomes
US6059724A (en) * 1997-02-14 2000-05-09 Biosignal, Inc. System for predicting future health
US6110109A (en) * 1999-03-26 2000-08-29 Biosignia, Inc. System and method for predicting disease onset
US20040265874A1 (en) * 2000-10-17 2004-12-30 Bio-Rad Laboratories, Inc. Pattern recognition method for diagnosis of systemic autoimmune diseases
US7640175B1 (en) * 2000-12-08 2009-12-29 Ingenix, Inc. Method for high-risk member identification
JP4389585B2 (en) 2001-10-19 2009-12-24 旭硝子株式会社 Substrate with transparent conductive oxide film and photoelectric conversion element
US20040030584A1 (en) * 2002-08-12 2004-02-12 Harris Jeffrey Saul System and method for guideline-based, rules assisted medical and disability management
US8560582B2 (en) * 2002-08-12 2013-10-15 Jeffrey Saul Harris Method for analyzing records in a data base
CN100444793C (en) * 2003-09-11 2008-12-24 株式会社日立医药 Organism light measuring device
WO2007084977A2 (en) * 2006-01-20 2007-07-26 Beckman Coulter, Inc. Low hemoglobin concentration cell percentage and method of use in detection of iron deficiency
WO2010053743A1 (en) * 2008-10-29 2010-05-14 The Regents Of The University Of Colorado Long term active learning from large continually changing data sets
US11857293B2 (en) * 2008-10-29 2024-01-02 Flashback Technologies, Inc. Rapid detection of bleeding before, during, and after fluid resuscitation
US20110172545A1 (en) * 2008-10-29 2011-07-14 Gregory Zlatko Grudic Active Physical Perturbations to Enhance Intelligent Medical Monitoring
US11395594B2 (en) * 2008-10-29 2022-07-26 Flashback Technologies, Inc. Noninvasive monitoring for fluid resuscitation
US11406269B2 (en) * 2008-10-29 2022-08-09 Flashback Technologies, Inc. Rapid detection of bleeding following injury
US11478190B2 (en) * 2008-10-29 2022-10-25 Flashback Technologies, Inc. Noninvasive hydration monitoring
US8512260B2 (en) 2008-10-29 2013-08-20 The Regents Of The University Of Colorado, A Body Corporate Statistical, noninvasive measurement of intracranial pressure
US11395634B2 (en) * 2008-10-29 2022-07-26 Flashback Technologies, Inc. Estimating physiological states based on changes in CRI
US11382571B2 (en) * 2008-10-29 2022-07-12 Flashback Technologies, Inc. Noninvasive predictive and/or estimative blood pressure monitoring
US20160038043A1 (en) * 2008-10-29 2016-02-11 Flashback Technologies, Inc. Assessing Effectiveness of CPR
US20100204590A1 (en) * 2009-02-09 2010-08-12 Edwards Lifesciences Corporation Detection of Vascular Conditions Using Arterial Pressure Waveform Data
JP5538749B2 (en) * 2009-06-03 2014-07-02 キヤノン株式会社 Diagnosis support system, diagnosis support method and program
US9938557B2 (en) 2010-09-16 2018-04-10 The General Hospital Corporation Red blood cell dynamics for administering treatment for iron-deficiency anemia
CA2871608C (en) * 2011-07-22 2020-07-07 Flashback Technologies, Inc. Hemodynamic reserve monitor and hemodialysis control
CN105389465B (en) * 2015-10-29 2018-09-28 徐翼 Medical record data processing method, device and system
US10955423B2 (en) 2015-12-15 2021-03-23 The General Hospital Corporation Methods of estimating blood glucose and related systems
US11293852B2 (en) 2016-04-07 2022-04-05 The General Hospital Corporation White blood cell population dynamics
US11918386B2 (en) 2018-12-26 2024-03-05 Flashback Technologies, Inc. Device-based maneuver and activity state-based physiologic status monitoring

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
SU1238105A1 (en) * 1984-12-05 1986-06-15 Военно-Медицинская Ордена Ленина Краснознаменная Академия Им.С.М.Кирова Computing device for disease diagnosis
US4844086A (en) * 1987-04-13 1989-07-04 The Children's Medical Center Corporation Cross correlation analysis in brain electrical activity mapping
US5083571A (en) * 1988-04-18 1992-01-28 New York University Use of brain electrophysiological quantitative data to classify and subtype an individual into diagnostic categories by discriminant and cluster analysis
EP0354716A1 (en) * 1988-08-03 1990-02-14 Chevron Research And Technology Company Apparatus and method for avoiding a drill string becoming stuck during drilling
WO1990008325A1 (en) * 1989-01-17 1990-07-26 Macri James N Down syndrome screening method
US5012675A (en) * 1989-07-25 1991-05-07 Amoco Corporation Integrating multiple mappable variables for oil and gas exploration

Also Published As

Publication number Publication date
DE69432885D1 (en) 2003-08-07
CN1121207A (en) 1996-04-24
EP0646881A2 (en) 1995-04-05
US5619990A (en) 1997-04-15
JPH07105166A (en) 1995-04-21
KR950009256A (en) 1995-04-21
EP0646881B1 (en) 2003-07-02
EP0646881A3 (en) 1996-03-06
TW290663B (en) 1996-11-11
KR100328119B1 (en) 2002-06-22
CN1099652C (en) 2003-01-22

Similar Documents

Publication Publication Date Title
JP3703858B2 (en) Attribution discrimination device
CN109036553B (en) Disease prediction method based on automatic extraction of medical expert knowledge
US10722179B2 (en) Residual-based monitoring of human health
CN104809226B (en) A kind of method of the uneven multivariate time series data of early stage classification
Starfield Measurement of outcome: a proposed scheme
Liu et al. Development and validation of a machine learning algorithm and hybrid system to predict the need for life-saving interventions in trauma patients
CN110051324B (en) Method and system for predicting death rate of acute respiratory distress syndrome
CN108135548A (en) For monitoring the method and system of pressure state
Ho et al. Imputation-enhanced prediction of septic shock in ICU patients
Hand Statistical methods in diagnosis
Amaral et al. Differential diagnosis of asthma and restrictive respiratory diseases by combining forced oscillation measurements, machine learning and neuro-fuzzy classifiers
Young et al. Interactive scaling with individual subjects
Fleiss et al. On the use of inverted factor analysis for generating typologies.
Green Evaluating the discriminatory power of a multiple logistic regression model
AU2018200093A1 (en) Residual-based monitoring of human health
Jørgensen et al. Use of neural networks to diagnose acute myocardial infarction. I. Methodology
Derevyagin et al. Applying Neural Networks in Polygraph Testing
Koukouvinos et al. Classification methods and ROC analysis for outcome prediction of patients following injuries
Power et al. Separation of normals, neurotics and simulating malingerers on the MPI by means of discriminant function analysis
CN115620917A (en) Intensive care data processing method, application and equipment
McGonigal A new technique for survival prediction in trauma care using a neural network
EISEN et al. Defining measurement precision for effort dependent tests: the case of three neurobehavioural tests
Trace et al. Testing pattern recognition as a method for measuring severity of illness
Harris et al. The preliminary credibility assessment system embedded algorithm description and validation results
Ballante Statistical and Machine Learning Models for Neurosciences

Legal Events

Date Code Title Description
A02 Decision of refusal

Free format text: JAPANESE INTERMEDIATE CODE: A02

Effective date: 20040113

A521 Written amendment

Free format text: JAPANESE INTERMEDIATE CODE: A523

Effective date: 20040312

A911 Transfer to examiner for re-examination before appeal (zenchi)

Free format text: JAPANESE INTERMEDIATE CODE: A911

Effective date: 20040428

A912 Re-examination (zenchi) completed and case transferred to appeal board

Free format text: JAPANESE INTERMEDIATE CODE: A912

Effective date: 20040604

A61 First payment of annual fees (during grant procedure)

Free format text: JAPANESE INTERMEDIATE CODE: A61

Effective date: 20050721

R150 Certificate of patent or registration of utility model

Free format text: JAPANESE INTERMEDIATE CODE: R150

LAPS Cancellation because of no payment of annual fees